AI Reference Finder — Real, Verifiable DOIs for Biomedical Claims

The hallucinated-citation problem is real

When researchers first used ChatGPT and other general-purpose LLMs to find references for medical and scientific claims, a failure mode emerged quickly: the model would return fluent, plausible, authoritative-looking citations that did not exist. Peer-reviewed analyses documented the scale of the problem:

• Bhattacharyya and colleagues (2023, Cureus) examined 178 references generated by ChatGPT in response to medical queries and found that 69 were fabricated entirely — the papers did not exist — and a further substantial fraction contained inaccurate DOIs or author lists.¹
• Later analyses across JAMA-family journals, radiology, urology, and psychiatry reproduced the pattern: double-digit percentages of LLM-generated medical citations were either fabricated or mis-attributed.

For any output that feeds a manuscript, a grant, a thesis, or a clinical decision, an unverifiable reference is a liability — it can trigger desk rejection at submission, correction notices at publication, or worse. An AI reference finder that cannot guarantee real DOIs is not fit for research use.

How BioSkepsis returns real references

BioSkepsis is built around a simple invariant: if a reference appears in the output, it exists in the corpus. Four mechanisms enforce that:

1. Retrieval-first, generation-second

Most hallucinated citations come from models that generate a reference string directly. BioSkepsis does the opposite. It retrieves candidate papers from its curated 40M+ biomedical corpus first, then uses the LLM layer only to rank, explain, and quote what retrieval returned. There is no step at which a reference is invented from training-data memory.

2. Every citation carries a verifiable DOI

Each returned reference ships with its DOI, PMID (where applicable), journal, year, and authors — all sourced from the canonical record, not regenerated. Click the DOI; it resolves to the publisher's landing page.

3. Biology-native disambiguation

Gene symbols, MeSH descriptors, and Gene Ontology terms are resolved to canonical biological entities before retrieval runs. A query mentioning "BRCA1" does not match papers that happen to use the string in an unrelated context; a query about "lupus" includes papers indexed under SLE even if the string "lupus" does not appear.

4. Declines when evidence is insufficient

If no paper in the corpus adequately supports a claim, BioSkepsis says so rather than returning a weak or off-topic reference. This is the single most important behaviour to check in any AI citation tool.

Use cases

Zotero and reference-manager export

BioSkepsis supports one-click export to the major reference managers:

• Zotero — direct push to your library, with collections preserved.
• Mendeley — BibTeX/RIS export, ready to import.
• EndNote — RIS export.
• LaTeX users — BibTeX export with canonical citation keys.

Every exported entry carries DOI, PMID, authors, journal, year, volume, issue, and pages — the fields you need for a conforming bibliography in any major style (APA, Vancouver, AMA, Harvard, MLA, Chicago).

BioSkepsis vs generic LLM citation generators

The takeaway: a generic artificial intelligence citation generator can be useful for brainstorming what a reference list might look like. An AI reference finder built for research — like BioSkepsis — is what you use when the references have to be real.

Frequently asked questions